# Executive summary
---
The two core storage mechanisms and the Table 1 byte accounting reproduce with the official [databricks/flashoptim](https://github.com/databricks/flashoptim) CUDA implementation: live optimizer state used ULP correction plus companded 8-bit moments, and measured storage was 7.125 bytes/parameter (5.125 with executed gradient release), including 0.125 bytes of FP16 group scales omitted by the table convention. ULP reconstruction and companding error advantages also reproduced, while the paper's 8B Llama peak-memory, 10B-token GPT-2, ImageNet ResNet-50, and end-task score suites were not rerun under the one-hour cap. A three-seed 300-step proxy closely tracked FP32 AdamW (Pearson r 0.999974–0.999994), but is labeled a toy proxy rather than evidence for those named end tasks. The registered local audits took about 20 seconds on one RTX PRO 6000; the larger public [Tesla T4 Job](https://huggingface.co/jobs/Umong/6a5a2040bee6ee1cf4ecd53b) took 14.25 seconds and cost well below $0.01.
## Scope & cost
| | This reproduction | Full replication |
| --- | --- | --- |
| Scope | Official representations, direct byte accounting, ULP/quantization error, toy convergence | Llama-3.1-8B finetuning, GPT-2 on 10B tokens, ResNet-50 on ImageNet, 3 seeds |
| Hardware | 1× RTX PRO 6000 locally; 1× Tesla T4 hosted | Distributed NVIDIA H100 GPUs |
| Compute time | ~20 s local; 14.25 s hosted | Multi-run, large-model training; far above one hour |
| Cost | Local sunk cost; hosted < $0.01 | Not estimated; substantial |
| Outcome | Claims 1–2 full; 3–5 source/partial; 6 partial | Not attempted |
---
````html
